MCP Servers
A collection of standalone Model Context Protocol
servers written in Go. Each server is independently buildable and installable — pick
only what you need. Released under the MIT No Attribution License.
Available Servers
| Server | Description | CGO |
|---|---|---|
| markitdown | Convert documents to Markdown (HTML, PDF, DOCX, XLSX, PPTX, images) | No |
| kg | Project knowledge graph — store and query code entities across sessions. Supports multi-scope graphs for monorepos, federated across layers, plus a personal store. Renders as mermaid or Graphviz diagrams with kg graph. Graphs can also be pushed to a shared hub (kg hub serve) and searched from other projects as remote layers. |
Yes |
Plus one shared library, not installed as a binary:
| Library | Description |
|---|---|
| kglib | KuzuDB schema, entity/relation/observation CRUD, hybrid search, embedders, and the FederatedStore that merges results across multiple KG databases |
graph LR
Client["MCP Client\n(Claude Code / Claude Desktop)"]
MD["markitdown-mcp\nDocument converter"]
KG["kg\nKnowledge graph"]
KGDB[(".ai/knowledge.db\nper project")]
PERSONAL[("~/.kg/knowledge.db\nper user")]
HUB["kg hub serve\nshared hub, per org"]
Client -->|stdio| MD
Client -->|stdio| KG
KG --> KGDB
KG -->|"--personal / --with-personal"| PERSONAL
KG -->|"HTTP: kg push / remotes"| HUB
Each server is a self-contained Go binary with its own go.mod. No server depends on
another; kg uses the kglib library via a local replace directive.
Quick Install```bash
Install all servers
curl -fsSL https://raw.githubusercontent.com/Cortexa-LLC/mcp/main/install.py | python3
Install a specific server
curl -fsSL https://raw.githubusercontent.com/Cortexa-LLC/mcp/main/install.py | python3 - --mcp kg
curl -fsSL https://raw.githubusercontent.com/Cortexa-LLC/mcp/main/install.py | python3 - --mcp markitdown
From a clone
python3 install.py --mcp kg
python3 install.py --mcp markitdown
python3 install.py # installs all
**Install dir defaults** (override with `--prefix DIR` or `INSTALL_DIR=...`):
| Platform | Default |
|----------|---------|
| macOS | `/usr/local/bin` |
| Linux | `/usr/local/bin` (may need `sudo`) |
| Windows | `%LOCALAPPDATA%\Programs\mcp` |
## Manual Build (per server)```bash
cd src/markitdown && make install
cd src/kg && make install # requires a C compiler (CGO)
make install # from the repo root: both servers
make test # kglib + all servers
Prerequisites
- Go 1.24+ — install
- kg only: C compiler — Xcode CLT on macOS (
xcode-select --install), gcc/clang on Linux - markitdown OCR
(optional): Tesseract 5+ (
brew install tesseract)
MCP Configuration
After installing, add the servers to your MCP client:
Claude Desktop```json
{
"mcpServers": {
"markitdown": {
"command": "/usr/local/bin/markitdown-mcp"
},
"kg": {
"command": "/usr/local/bin/kg",
"args": ["server", "--stdio"]
}
}
}
Config file locations:
- **macOS**: `~/Library/Application Support/Claude/claude_desktop_config.json`
- **Linux**: `~/.config/Claude/claude_desktop_config.json`
- **Windows**: `%APPDATA%\Claude\claude_desktop_config.json`
### Claude Code (`.mcp.json` in project root)
**Recommended for `kg`** — place `.mcp.json` in each project root so Claude Code
spawns `kg` with the correct working directory, ensuring it opens that project's
`.ai/knowledge.db` and not another project's graph.```json
{
"mcpServers": {
"markitdown": { "command": "/usr/local/bin/markitdown-mcp" },
"kg": { "command": "/usr/local/bin/kg", "args": ["server", "--stdio"] }
}
}
markitdown is stateless and can be configured globally in ~/.claude/settings.json
if preferred.
Server Details
-
markitdown — No CGO, no system dependencies for core formats.
OCR for images requires Tesseract (optional, degrades gracefully).
→ Integration guide -
kg — Requires CGO (bundles KuzuDB statically). Each project gets
its own isolated graph at.ai/knowledge.db, auto-discovered by walking up the directory
tree. Monorepos can split the graph into scopes that federate across layers. Supports
OpenAI or Ollama embeddings for semantic search (optional).
Also keeps a personal store (kg personal init) for knowledge that follows you
between repositories, searchable alongside a project's graph with--with-personalor via
thesearch_personal_knowledgeMCP tool. Agents can record to it only when explicitly
enabled with--personal-writes. -
kglib — Shared library, not a server. Owns the schema,
search, embedders, Cypher read-only guard, andFederatedStore. Read this before
changing storage or federation behaviour in either graph.
Docs
| Document | Description |
|---|---|
| docs/kg-cli-reference.md | Full kg CLI reference — all commands, flags, entity types, Cypher examples |
| docs/kg-scopes.md | Multi-scope monorepo graphs and federated (layered) search — configuration and commands |
| docs/kg-scopes-implementation.md | How scopes and federation are implemented internally |
| docs/kg-graph-linking-design.md | Design and measurement behind cross-layer entity linking in kg graph --federated — why the old (name, type) join was wrong and what replaced it |
| docs/kg-personal-store.md | The personal knowledge store — creating it, writing to it, and federating it into project searches |
| docs/kg-shared-service.md | The shared knowledge hub — running it, seeding it with kg push, and searching hub graphs from other projects via remotes |
| docs/kg-shared-service-design.md | The hub's original design proposal — architecture, threat model, and hosting options |
| src/kglib/README.md | kglib library API, including FederatedStore merge semantics |
| skills/README.md | Claude Code skills built on these servers, and how to install them |
| docs/kg-log-plugins.md | Log-file indexing plugins — how a plugin is matched to a log format, and how to add one |
| docs/kg-claude-integration.md | KG patterns for CLAUDE.md, reducing re-investigation, decision logging, cross-session checkpointing |
| docs/markitdown-claude-integration.md | Reading PDFs, DOCX, spreadsheets, and URLs; combining with the KG |
| CHANGELOG.md | Notable changes since v0.1.0 |
